Forward Future by Matthew Berman

On-Device AI, OpenAI Sol Safety, & GPT-5.6 Efficiency

Matthew Berman — Fri, 10 Jul 2026 12:00:00 +0000

Connect Higgsfield to your Claude for free — takes about a minute

Good morning. It's Friday, July 10, and we're covering AI's push onto smartphones, questions over who approves frontier AI models, and why OpenAI is betting on efficiency over raw intelligence.

Plus: yesterday's poll results on AI for planned economies, and our latest Zeitgeist discussion on AI's shift toward price-performance.

YOUR DAILY ROLLUP

One Second of Your Mind

In 2013, one of the fastest machines on Earth tried to keep pace with a human brain for a single second, and the result says a lot about the hardware sitting inside your skull. Get the full picture at the bottom of today's issue.

POWERED BY HIGGSFIELD AI

Claude Just Learned to Make Videos

Higgsfield now plugs straight into Claude via MCP. Ask for a cinematic ad, a talking avatar, a product video, and it renders inside the chat while you keep working. No tabs, no timeline, no learning curve. The whole studio, one prompt away.

Connect Higgsfield to your Claude for free — takes about a minute

VIDEO

GPT 5.6 Becomes a Workhorse

GPT 5.6 squeezes more from the same model, building complex software, excelling at computer use, and cutting inference costs.

FORWARD FUTURE REVIEWS

Matt Spent 25 Billion Tokens on GPT-5.6 — Here's His Verdict

GPT-5.6 is here after a two-week government delay, and Matt has been testing it internally for two months. The full review covers what nobody else has yet: a Minecraft-style game the model kept building for days after he walked away, an Excel clone it refined over six days using Computer Use, and browser control fast enough to manage live Supabase instances unprompted.

Plus the head-to-head everyone wants: GPT-5.6 vs. Claude Fable, from benchmarks to cost per finished task. Every demo in the review is live and playable.

Read the full review →

ON-DEVICE AI

PrismML Claims Record 27B-Parameter AI Model Running on iPhone

PrismML says it has compressed Alibaba’s 27-billion-parameter Qwen 3.6 open-source AI model to run entirely on an iPhone 17 Pro, making it the largest AI model yet demonstrated on a smartphone. The startup claims its compression technique reduces the model from about 54GB to under 4GB without degrading performance, enabling advanced capabilities such as reasoning, coding, autonomous agents, and complex chat directly on-device.

The breakthrough aligns with Apple's push to move more AI processing onto iPhones to improve privacy and reduce cloud infrastructure costs, and Apple has reportedly held discussions with PrismML about its technology. While the claims have not been independently verified, the company plans to release the compressed model next week and says it intends to apply the same approach to even larger frontier-scale AI models. → Read the full article here. (Paywall)

REGULATION

Questions Persist Over Government Approval of OpenAI's Sol AI Model

OpenAI has begun rolling out its frontier AI model, Sol, but the process the U.S. government used to determine it was safe for public release remains largely opaque. Researchers, policymakers, and industry leaders told TechCrunch they lack visibility into the evaluation criteria, who conducted the assessments, or which standards models must meet before deployment. OpenAI says Sol underwent external safety testing by organizations including the U.K.

AI Security Institute (AISI), SecureBio, and Irregular, while CEO Sam Altman confirmed discussions with senior U.S. officials, though neither the company nor the government has disclosed details of the approval process. The uncertainty has fueled concerns that frontier AI oversight relies on informal government-industry relationships rather than a transparent, standardized framework, raising questions about accountability as increasingly powerful models reach the public. → Read the full article here.

ZEITGEIST

The Week AI Became a Cost Question

Every morning the Forward Future team argues about the stories, papers, and launches that matter. What's real, what's hype, and where AI is headed next. These are the public notes.

Are frontier AI models still the ones that matter, or are cheaper, "good enough" models quietly taking over? The debate ranged from benchmark obsession and open-source adoption to the growing gap between pre-release demos and public launches.

▸ Why price-performance is beating raw intelligence
▸ Could open source dominate AI inference by year-end?
▸ The growing credibility gap between early demos and public releases

Read yesterday's Zeitgeist →

NEWS

What Else is Happening

OpenAI Kills Atlas: OpenAI is shutting down Atlas less than a year after launch and redistributing its agentic features into a ChatGPT Chrome extension.

AI IPOs Eclipse Tech History: SpaceX, Anthropic and OpenAI could surpass the combined value of all U.S. VC-backed exits since 2000.

Meta’s In-House AI Chip Strategy: Meta moves its custom AI chip effort into production as it expands infrastructure beyond third-party hardware.

Palo Alto Networks CEO Targets AI Costs: CEO Nikesh Arora says AI token costs must fall 90% within two years to drive broader enterprise adoption.

Character.AI Debuts Interactive Microdramas: New AI microdramas let adult viewers chat and roleplay with characters, expanding its entertainment push.

China Flags Claude Code Flaws (Paywall): China says Anthropic's Claude Code exposed user data through built-in monitoring, urging updates or removal.

FRIDAY FACT

One Second of Brain Activity Took 40 Minutes To Simulate

In 2013, researchers used Japan's K computer — then among the fastest in the world — to model the activity of a neural network for one second. The machine ran 82,944 processors and about a petabyte of memory to simulate 1.73 billion neurons connected by 10.4 trillion synapses. It took 40 minutes of computing to reproduce a single second. And that network represented roughly one percent of the human brain.

Your brain does the full version continuously, on about 20 watts — less power than the bulb in your refrigerator. The 2013 run wasn't trying to explain how the mind works; the neurons were wired at random. It was a stress test of the software and the hardware. But the gap it exposed is the point. The most efficient computer we know of is the one that reads this sentence, and we still can't simulate a hundredth of it in real time.

TWEETS

NEO’s Hands To Tackle Human Tasks

Twitter tweet

POLL RESULTS

Does AI Make Planned Economies Viable?

Here’s how you voted: Nearly half said AI doesn't fix the core problems of planned economies, while 32% warned it could create a bigger monopoly.

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

AI Search Debate, Infrastructure Debt Boom, & Grok 4.5

Matthew Berman — Thu, 09 Jul 2026 16:03:24 +0000

Connect Higgsfield to your Claude for free — takes about a minute

Good morning. It's Thursday, July 9, and we're covering whether AI search is making us less curious, Wall Street's AI infrastructure spending spree, and Grok 4.5's push for faster, cheaper AI.

Plus: yesterday's Real or AI verdict and a new poll asking whether AI changes the case for planned economies.

YOUR DAILY ROLLUP

Claude Just Learned to Make Videos

Connect Higgsfield to your Claude for free — takes about a minute

VIDEO

The Hidden Mind of AI

Anthropic's new JSpace research reveals how AI reasons internally, offering unprecedented insight into Claude's hidden thought process.

ZEITGEIST

When AI Knows Too Much

Every morning the Forward Future team argues about the stories, papers, and launches that matter. What's real, what's hype, and where AI is headed next. These are the public notes.

Yesterday's fight: what happens when better models make information cheap, abundant, and hard to protect? The conversation moved from data moats and open-source economics to energy demand and voice interfaces.

▸ Data moats in a world of shared intelligence
▸ Jevons, energy, and the demand curve for intelligence
▸ Open source is not replacing frontier models yet

Read yesterday’s Zeitgeist →

POLL

Does AI Make Planned Economies Viable?

One of the many knocks on communism: no central planner can distribute resources as efficiently as free markets. But what if AI can? If a superintelligence has access to perfect information and can allocate resources better than the price mechanism.

CURIOSITY

AI Search May Be Making Us Less Curious, Neuroscientist Argues

In an opinion essay, neuroscientist Anne-Laure Le Cunff argues that AI-powered search tools such as Google’s AI Overviews, ChatGPT, Claude, and Perplexity are reducing opportunities for curiosity by delivering instant answers instead of encouraging exploration. Drawing on neuroscience research, she says curiosity strengthens learning when people spend time seeking answers, allowing the brain to absorb related information and form broader connections.

Le Cunff contends that AI-generated summaries shorten this process, limiting incidental discoveries that have historically fueled scientific, technological, and creative breakthroughs. She calls on AI companies to redesign search experiences by making sources more visible, presenting competing perspectives, and creating modes that prioritize exploration alongside efficiency. → Read the full article here. (Paywall)

INVESTMENT

AI Infrastructure Boom Fuels Record Corporate Bond Issuance

Technology companies are raising unprecedented amounts of debt to finance AI infrastructure, extending the AI investment boom from equity markets into corporate bonds. Since January 2026, Meta, NVIDIA, Oracle, SpaceX, Amazon, and Alphabet have collectively issued tens of billions of dollars in bonds, with Morgan Stanley forecasting $350–400 billion in AI-related investment-grade issuance in the U.S. this year.

While highly rated "hyperscalers" have kept borrowing costs low, investors are becoming more cautious about riskier AI-linked projects such as data centers and cloud infrastructure providers. The article argues that assessing long-term credit risk is increasingly difficult as the AI industry evolves rapidly, making it harder to predict which companies will generate enough returns to repay debt over the coming decades. → Read the full article here. (Paywall)

NEWS

What Else is Happening

Prime Intellect Raises $130M: Startup hits a $1B valuation after funding to expand its platform helping enterprises build and train custom AI agents.

OpenAI Unveils GPT-Live Voice Model: Instead of waiting for users to finish speaking, GPT-Live can listen and respond simultaneously.

NVIDIA Loses $1T in Value: Shares fell 16% from May highs, leaving the AI chip leader trading at its lowest valuation since before the AI boom.

BofA Lends OpenAI $520M: Bank of America extended its first credit line to OpenAI as the AI company prepares for a potential IPO.

SambaNova Raises $1B: AI chip startup reached an $11B valuation after new funding to expand inference hardware and on-premise AI systems.

POLL RESULTS: REAL OR AI

Can You Still Tell if an Image Is AI?

Yesterday's image was REAL. Here's how you voted: 69% of readers thought it was AI, while 31% correctly identified the image as real.

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

Chinese AI Models, Illinois AI Law, & Alibaba Bans Anthropic

Matthew Berman — Wed, 08 Jul 2026 12:00:00 +0000

Good morning. It's Wednesday, July 8, and we’re covering why U.S. companies are turning to Chinese AI models, Illinois' landmark AI safety law, and Alibaba's ban on Anthropic.

Plus: yesterday' poll results on AI oversight, and a guest post on why mixed coding agents create a new challenge for engineering teams.

YOUR DAILY ROLLUP

Slash AI Costs With Model Routing

Learn how model routing cuts AI coding costs by pairing premium models for planning with lower-cost models for execution.

FORWARD FUTURE ORIGINAL

The New Coding Agents Problem: They Don’t Work Together

The past year saw an explosion of coding agents, like Claude Code, Codex, Gemini, and Warp Agent, highly capable tools increasingly taking on the work of a junior engineer. Developers have spent the year testing and trialing them, and individuals have gravitated toward one or another, not necessarily because it performs better or is more capable, but because of personal preference.

As a result, software engineering teams are growing fleets of mixed agents: some people using one, some using another. → Read the full article here.

COMPETITION

Chinese AI Models Gain Ground as U.S. Companies Cut AI Costs

U.S. companies are increasingly adopting AI models from Chinese developers such as DeepSeek and Z.ai as rising costs from leading U.S. providers, including OpenAI and Anthropic, push businesses to seek more affordable alternatives. Data from OpenRouter shows Chinese models have accounted for more than 30% of weekly token usage by U.S. companies since February 8, 2026, with some platforms reporting rapid adoption driven by prices that are 60% to 90% lower than comparable U.S. models.

Companies including Lindy say switching to Chinese models has reduced AI expenses dramatically while maintaining or even improving performance for many workloads. As Chinese models continue to narrow the capability gap with leading American systems, businesses are increasingly choosing models based on cost-effectiveness rather than brand, reshaping competition in the AI market amid growing U.S. regulatory scrutiny. → Read the full article here.

REGULATION

Illinois Enacts AI Safety Law Targeting Largest AI Model Developers

Illinois Gov. JB Pritzker signed the Artificial Intelligence Safety Measures Act on July 7, 2026, establishing new transparency, reporting, and audit requirements for the largest AI model developers. The law applies to models generating more than $500 million in annual revenue and requires companies to assess catastrophic risks, report serious AI-related incidents within 72 hours—or 24 hours if an imminent threat exists—and undergo annual independent audits.

Modeled on AI safety laws passed in California and New York, the legislation is intended to create consistent guardrails for advanced AI systems while Congress has yet to adopt a federal framework. The law passed with broad bipartisan support, was backed by OpenAI and Anthropic, and will take effect on January 1, 2028. → Read the full article here.

ROBOTICS

AI Advances Bring Autonomous Robots Closer to Everyday Work

Advances in AI are accelerating the development of autonomous robots capable of performing increasingly complex tasks in warehouses, factories, and other real-world environments, though experts say truly general-purpose robots remain years away. Companies including Boston Dynamics, Agility Robotics, and Physical Intelligence are combining reinforcement learning, foundation models, and larger training datasets to improve robots' ability to navigate unpredictable settings and complete multi-step tasks with minimal human supervision.

Humanoid robots are already handling logistics and manufacturing work, while developers continue to prioritize safety through new standards, controlled deployments, and AI-assisted decision-making before expanding into human-centric environments. Researchers expect workplace adoption to grow steadily, but caution that reliable robots for homes and fully autonomous surgical systems will require significant technological and safety advances before becoming mainstream. → Read the full article here.

NEWS

What Else is Happening

AI Actor Tilly Norwood Debuts: Particle6 cast its AI-generated performer in feature film Misaligned, reviving debate over AI's role in filmmaking.

Claude Cowork Goes Mobile: Anthropic brought Claude Cowork to web and mobile for Max users, enabling AI tasks to run across devices.

Microsoft Expands In-House AI: Microsoft is replacing some OpenAI and Anthropic models in Excel and Outlook with MAI to cut AI costs.

Samsung Slides Despite Profit Surge: Samsung shares fell after a 19-fold profit jump as investors took AI chip profits, dragging memory stocks lower.

Study Warns on AI Finance Advice: Researchers found leading AI chatbots gave inconsistent, inaccurate or biased personal finance recommendations.

Amazon AI Bond Demand Cools: Amazon's $25 billion bond sale drew weaker demand as investors grew more selective amid rising AI-related debt.

REAL OR AI

Can You Still Tell if an Image Is AI?

POLL RESULTS

Can AI Be Trusted Unsupervised?

Here’s how you voted: The poll was split, with 41% alarmed, 41% seeing human-like behavior, and 18% reassured.

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

The New Coding Agents Problem: They Don’t Work Together

Zach Lloyd — Tue, 07 Jul 2026 23:35:50 +0000

_{Zach Lloyd, Founder and CEO of Warp, an AI-native developer platform. Previously a Principal Engineer at Google, he led engineering for Google Docs and Google Sheets.}

This isn't a new phenomenon: it’s how developers have always worked. Nobody expects every engineer to use the same code editor, terminal, or to organize their workflow the same way, and most organizations are set up structurally to support a variety of tools being used across the team.

Developers have always picked their own tools, and that's unlikely to change any time soon. When it comes to coding agents, though, that freedom of choice presents logistical challenges. Managing fleets of mixed coding agents is inherently difficult, and as agents move off of developer’s local machines and into the cloud, teams will require new infrastructure to support both centralized agentic systems and developer choice.

What Breaks When Agents Don’t All Come From the Same Place

Agents from different providers weren't designed to work together. You may have already seen this play out in unexpected ways, like reports of OpenAI models making OpenClaw suddenly obsessed with goblins. It's funny, but it's also a glimpse of what happens when two tools from different providers don't quite speak the same language.

The most immediate problem is visibility. When agents come from different providers, each one surfaces what it's doing differently: different logging formats, different interfaces, different levels of transparency. There's no unified view of what ran, what it touched, or where something went wrong. Debugging a multi-agent workflow today often means stitching together context from three different dashboards that have nothing in common.

Permissions are also an issue. An agent working on deployment infrastructure shouldn't have the same access as one summarizing logs, but managing secrets, API keys, and environment settings across different agent harnesses adds real overhead. Most teams end up either over-permissioning for convenience or managing access manually in ways that don't scale.

Memory is a third problem. Most agents forget everything between sessions—a limitation even within a single harness. Across providers, there's no shared context at all by default. Every new session starts from zero. A team that worked through a complex refactor last week has to re-explain the approach, the constraints, and the decisions already made. That's not just inefficient, it multiplies across every session and every engineer on the team.

Cost compounds all of this. When developers choose tools on preference and stick with them, there's no natural mechanism pushing toward the most cost-efficient option. With past dev tools, this hasn’t been a major problem because SaaS subscriptions are easy to forecast. AI token costs, however, can quickly spiral, and engineering leaders are increasingly asking how to get visibility into what they're actually spending across a fleet they don't fully control. Right now, most don't have a good answer.

What You’re Gaining From Mixed Fleets

None of these challenges mean the mixed fleet is working against you. When your team is managing them properly, the benefits go beyond having your preferred agent for each job. For one, you're not dependent on any single provider, so if one model is down, rate-limited, or underperforming on a specific task, you can route around it. That optionality matters more as agents become load-bearing infrastructure rather than productivity tools.

Cost is another factor. Routing more demanding tasks to one model and simpler tasks to cheaper alternatives sharply reduces cost without requiring you to rebuild your entire workflow. In a recent YC-Bench startup simulation, GLM-5 ran a simulated company about as well as Claude Opus, but at roughly one-tenth of the inference cost. Single-vendor stacks don't give you that lever, and they're not designed to. Every major provider is incentivized to keep you on their full stack. The agent that works best for your use case today may not be the one you want in six months, and if you're running multiple agents already, swapping one out doesn't mean starting over.

Getting multiple agents to work well together requires being deliberate about how you run them. The teams I see pulling ahead are spending less time on benchmarks and more time on the layer that makes managing these fleets of agents possible: how agents share context, how permissions get enforced across providers, and how you maintain visibility into what's actually running. That's the real work, and the teams focused on it now will have a meaningful head start.

Zach Lloyd is the founder and CEO of Warp, a platform for agentic development used by leading AI labs, Big Tech, and over half of the Fortune 500. Before founding Warp, he spent nearly a decade as Principal engineer on the Google Docs suite, helping scale products like Sheets to hundreds of millions of users.

AI Competition, Superintelligence, & the Future of Work

Matthew Berman — Tue, 07 Jul 2026 11:01:00 +0000

2026 State of AI in the Enterprise: Box’s Report Available Now!

Good morning. It's Tuesday, July 7, and we're covering China's use of distillation, researchers’ belief that progress is about to accelerate, and tech leaders shifting their views on AI’s job impact.

Plus: a poll on whether AI can be trusted unsupervised.

YOUR DAILY ROLLUP

Box Report: Governance Must Be Designed for AI

76% of organizations say current AI governance is slowing their ability to deploy agentic AI— 93% agree that better governance would help them move faster. That is not a contradiction

It is governance retrofitted from human workflows vs. governance designed for agents: permissions, visibility, auditability, and trusted-source controls built in.

Download the report to learn more!

POLL

Can AI Be Trusted Unsupervised?

Anthropic published research today showing they can read Claude's hidden thoughts, and found that Claude sometimes behaves ethically only because it suspects it's being tested. Block those "I'm being watched" thoughts, and it threatens blackmail.

VIDEO

Codex Clones Excel Autonomously

Codex autonomously cloned Excel over 12 days, producing a polished spreadsheet app with working formulas, formatting, and sorting tools.

GEOPOLITICS

American AI Firms Warn China Is Closing the Gap Through Distillation

U.S. AI companies including Anthropic and OpenAI are urging lawmakers to curb the use of AI "distillation," a technique they say Chinese firms use to replicate the behavior of proprietary models. In a June 10, 2026 letter to Senators Tim Scott and Elizabeth Warren, Anthropic accused Alibaba of accessing its systems through thousands of unauthorized accounts to collect training data, allegations Alibaba declined to comment on.

Distillation is a long-established AI training method originally developed at Google, and even U.S. companies have acknowledged using it, leaving its legal status unsettled. The debate has intensified as Chinese models such as Z.ai's GLM-5.2 approach the capabilities of leading American systems, fueling concerns over AI competition, export controls, and national security. → Read the full article here. (Paywall)

PREDICTIONS

Researcher Predicts Rapid Intelligence Gains and a Stranger Future

A widely shared essay titled "46 Thoughts on the Near Future" argues that AI capabilities are poised to accelerate much faster than governments, businesses, and the public expect, driven by algorithmic breakthroughs, AI-assisted research, and expanding compute. The author predicts that increasingly autonomous AI will transform scientific discovery, robotics, manufacturing, and the global economy while reshaping labor markets, potentially creating both unprecedented prosperity and new social inequalities.

The essay also warns that faster AI progress could heighten cybersecurity, military, and governance risks, making international coordination and stronger institutions more important. While presented as informed forecasts rather than established facts, the piece reflects a growing view among some AI researchers that the coming decade could bring changes far beyond incremental improvements. → Read the full article here.

WORKFORCE

Tech CEOs Shift Away From AI Job Apocalypse Narrative

After years of warning that AI could eliminate large numbers of jobs, several leading tech executives are now emphasizing its potential to augment workers rather than replace them. OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Meta CEO Mark Zuckerberg, Amazon founder Jeff Bezos, and Amazon CEO Andy Jassy have all recently argued that AI can boost productivity, create new roles, and enable companies to accomplish more with existing teams, even as some continue to acknowledge the risk of long-term job displacement.

The shift comes despite ongoing layoffs and heavy AI investment, with new research suggesting companies adopting AI are often hiring more rather than less. Economists and business leaders remain divided, however, as real-world AI deployments have proven more gradual and complex than early predictions, leaving the technology's long-term impact on employment far from settled. → Read the full article here.

NEWS

What Else is Happening

Microsoft Cuts 4,800 Jobs: Microsoft is cutting 4,800 jobs as it restructures around AI, saying the roles aren't being replaced by AI.

NVIDIA Denies Kyber Delay: NVIDIA rejected claims its Kyber AI server is delayed to 2028, saying its roadmap remains intact despite challenges.

TeraWulf Lands $19B Anthropic Deal: TeraWulf signed a 20-year, $19B data center lease with Anthropic, deepening its shift from bitcoin mining.

Apple Adds Siri Voice Controls: iOS 27 beta 3 lets users adjust Siri's speaking pace and expressiveness for a more personalized AI experience.

TWEETS

The Smarter Way to Review Code

Twitter tweet

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

World Models, NVIDIA's Cloud Push, & Meta's AI Reset

Matthew Berman — Mon, 06 Jul 2026 11:01:00 +0000

Good morning. It's Monday, July 6, and we're covering why AI is moving beyond language models, how NVIDIA is fueling the AI infrastructure race, and why Meta says powerful AI agents are taking longer than expected.

Plus: a prompt to help you prepare for your next meeting.

YOUR DAILY ROLLUP

AI Is Driving Apple Prices

AI's memory boom is driving up RAM costs, which could make upcoming MacBooks and iPads significantly more expensive.

MODELS

AI Researchers Shift Focus From Language Models to World Models

AI pioneer Yann LeCun argues that today's large language models (LLMs), including ChatGPT, Claude, and Gemini, are powerful but fundamentally limited because they lack an understanding of the physical world. His Paris-based startup, AMI Labs, is developing a new approach called Joint Embedding Predictive Architecture (JEPA), backed by more than $1 billion in seed funding from investors including NVIDIA and Jeff Bezos' investment fund, with plans to begin industrial deployments in 2027.

Researchers such as Ingmar Posner believe the next wave of AI will come from "world models" that can reason about cause and effect, predict outcomes, and interact safely with real-world environments—capabilities seen as essential for robotics and more general-purpose artificial intelligence. → Read the full article here.

COMPETITION

NVIDIA Backs Neocloud Boom as AI Infrastructure Competition Intensifies

NVIDIA is expanding financial support for emerging "neocloud" providers—companies that rent AI computing infrastructure—as demand for AI chips continues to surge. The strategy includes revenue-sharing backstops that help smaller cloud operators secure financing, even as established providers such as Amazon Web Services, Google Cloud, Microsoft Azure, and CoreWeave rapidly expand capacity.

The move reflects NVIDIA's effort to sustain demand for its AI chips amid growing competition, but it also raises concerns that easy financing could fuel overbuilding and leave smaller startups vulnerable if AI infrastructure demand slows. The broader race underscores how access to compute has become one of the defining competitive battlegrounds in the AI industry. → Read the full article here. (Paywall)

BENCHMARKS

Meta Says New AI Model Has Reached GPT-5.5 Performance

Alexandr Wang told employees that Meta's next AI model, codenamed Watermelon, has matched the benchmark performance of OpenAI's GPT-5.5, according to people familiar with an internal town hall. Wang said the model is still in training and uses roughly ten times more computing power than Meta's current Muse Spark model, while highlighting upcoming improvements in coding and AI agent capabilities.

The claim has not been independently verified, and Meta has not publicly released benchmark results or commented on the comparison. If confirmed, it would signal that Meta's multibillion-dollar investments in AI talent, chips, and infrastructure are narrowing the performance gap with OpenAI, Google, and Anthropic. → Read the full article here.

NEWS

What Else is Happening

India probes Tata iPhone leak: Investigation follows a breach exposing unreleased iPhone 18 Pro files, supplier lists and component details.

Google Dials Back AI: Google's 250th-anniversary ad imagines the Founders drafting the Declaration in Docs and Meet, but keeps Gemini in the margins.

Tesla caps AI spending: Employees face a $200 weekly AI limit unless approved, while Grok beta use remains exempt under the new policy.

Argentina proposes AI-run companies: Milei's bill would recognize AI-operated firms but still requires human oversight and accountability.

Alibaba, Tencent back Kling AI: Kuaishou raised $2.8 billion for its AI video unit at a $15 billion valuation, with Alibaba, Tencent and Baidu taking stakes.

PROMPT OF THE WEEK

The Meeting Recall Prompt

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

AI Trading Agents, Economic Change, & OpenAI's Equity Plan

Matthew Berman — Fri, 03 Jul 2026 11:01:00 +0000

Good morning. It's Friday, July 3, and we're covering how AI agents are moving into investing, why AI's economic impact is proving difficult to track, and OpenAI's proposal to share AI wealth with the public.

Plus: yesterday's poll results on AI layoffs, and a Friday Fact on memory transfer.

YOUR DAILY ROLLUP

Can a Memory Be Injected Into Another Animal?

In 2018, a team of neuroscientists gave an animal a memory it never actually lived through — the full story is at the bottom of this newsletter.

VIDEO

The Problem With Staggered Releases

GPT-5.6's staggered rollout could concentrate AI access among a few companies, raising concerns about competition and government influence.

TRADING

Robinhood CEO Says AI Agents Will Soon Match Human Traders

Robinhood CEO Vlad Tenev said on July 2, 2026, that AI agents will soon be capable of performing the same trading tasks as human traders, extending sophisticated automated strategies beyond institutional firms to everyday investors. Robinhood introduced agentic AI tools in May that allow AI agents to trade stocks and make purchases on users’ behalf, part of a broader industry push led by companies including OpenAI and Anthropic.

Tenev said much of institutional trading is already automated and AI-powered, but these capabilities have historically been inaccessible to retail investors. The comments come as Robinhood expands into U.K. crypto trading and Europe, with the company serving nearly 28 million customers across 38 countries despite recent workforce cuts aimed at improving efficiency. → Read the full article here.

ECONOMICS

AI’s Economic Impact Is Growing Faster Than It Can Be Measured

An analysis by The New York Times highlights how artificial intelligence is spreading through the economy faster than economists and policymakers can reliably measure its effects. Current data paint conflicting pictures, with some studies linking AI to job losses—particularly among recent graduates—while others suggest the technology is boosting hiring, productivity, or helping ease inflationary pressures.

Researchers also disagree on fundamental questions, including how widely businesses have adopted AI and which workers are most at risk of disruption. The rapid adoption of generative AI in less than four years has outpaced traditional economic measurement, raising concerns that policymakers may struggle to respond before its full effects become clear. → Read the full article here. (Paywall)

REGULATION

The Economist Urges Smarter Rules Instead of Restricting Frontier AI

The Economist argues that the U.S. should establish clear, predictable regulations for frontier AI rather than relying on ad hoc restrictions that limit access to advanced models. The publication points to the Trump administration's temporary export controls on Anthropic's Fable model and subsequent policy reversal as evidence of an opaque regulatory approach that creates uncertainty for AI companies.

It argues that permanently restricting powerful AI models is both impractical—given rapid advances by Chinese developers—and counterproductive, since broad access helps economies adapt gradually to new capabilities. Instead, the editorial advocates for formal evaluations, phased releases to trusted institutions, and oversight by independent technical bodies working alongside governments. → Read the full article here. (Paywall)

NEWS

What Else is Happening

Electricity Demand Outpaces GDP: AI data centers and electrification will drive power demand above GDP growth, boosting demand for key industrial metals.

Singapore Seizes $42M Mansion: Authorities allege the luxury home was bought with proceeds from NVIDIA chip smuggling in a wider export probe.

Neo Takes on Microsoft Office: Bhavin Turakhia is investing $30M of his own money to build an AI-first workplace platform for businesses.

Anthropic Denies Government Stake Talks: A source says Anthropic has not discussed giving the U.S. government an ownership stake.

TWEETS

Fable 5 Sparks Performance Backlash

POLL RESULTS

The AI Layoff Reversal

Here's how you voted: The top answer was that AI couldn't do the job yet, followed closely by failed cost savings.

FRIDAY FACT

Scientists Transferred a Memory Between Two Snails

In 2018, a team led by David Glanzman at UCLA trained sea snails (Aplysia californica) by giving their tails a series of mild electric shocks, producing a heightened, long-lasting defensive withdrawal reflex.

The researchers then extracted RNA from the nervous systems of the trained snails and injected it into snails that had never been shocked. The untrained recipients immediately displayed the same sensitized reflex as the snails that had actually gone through the training. Snails injected with RNA from untrained donors showed no such change.

For over a century, memory has been understood as something stored in the strength of connections between neurons. This experiment, published in the journal eNeuro, suggests RNA itself can carry a memory-like instruction, likely through an epigenetic change called DNA methylation. It doesn't overturn the standard theory of memory, but it complicates the assumption that memories live only in synapses.

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

Vinton Cerf Retires, AI Policy Shifts, & Claude Fable 5 Returns

Matthew Berman — Thu, 02 Jul 2026 11:01:00 +0000

Good morning. It's Thursday, July 2, and we're covering the retirement of an internet legend, the AI rivalry between the U.S. and China, and Anthropic's latest effort to balance frontier AI with stronger safeguards.

Plus: yesterday's Real or AI verdict, and a new poll about the AI layoff reversal.

YOUR DAILY ROLLUP

The AI Layoff Reversal

Ford, IBM, Klarna, and Commonwealth Bank all laid off workers for AI, and are now quietly rehiring them. A new survey finds 32% of US hiring managers reversed an AI-driven layoff, and 55% of business leaders admit they made the wrong call.

VIDEO

Anthropic Alleges Distillation Attack

Anthropic alleges Alibaba used 25,000 fake accounts to distill Claude. Could KYC become mandatory for access to frontier AI models?

MARKET PULSE

Palo Alto & CrowdStrike Post Record Quarter on AI Cyber Demand

Palo Alto Networks and CrowdStrike posted their strongest quarter on record between April and June 2026, with shares rising 113% and 95%, respectively, as growing AI-powered cyber threats fueled demand for advanced security tools. The surge follows the emergence of powerful Mythos-class AI models, which have heightened concerns about automated vulnerability discovery and large-scale cyberattacks.

Both companies secured early access to Anthropic's Project Glasswing and OpenAI's Daybreak cybersecurity initiatives. They are also expanding investments in identity security and agentic AI defenses. CrowdStrike said Falcon Shield delivered fourfold annual recurring revenue growth, reinforcing expectations that AI will continue driving cybersecurity spending. → Continue reading here.

LEGACY

Internet Pioneer Vinton Cerf To Retire From Google After Two Decades

Vinton Cerf, widely known as the "Father of the Internet," will retire from his role as Google's chief internet evangelist in early July 2026 after more than 20 years with the company. Cerf, 83, co-developed the TCP/IP networking protocols with Robert Kahn in the 1970s, laying the foundation for the modern internet and earning honors including the Presidential Medal of Freedom and the Turing Award.

Speaking at the Open Frontier conference, Cerf argued that the rise of AI agents will require standardized communication protocols rather than relying solely on natural language, predicting interoperability will become essential as autonomous systems increasingly interact. → Read the full article here.

COMPETITION

Trump’s AI Policy Shifts Raise Concerns Over China’s Competitive Gains

The Trump administration’s shifting approach to advanced AI regulation has sparked concerns among security experts that inconsistent restrictions on U.S. AI models could give Chinese developers an opportunity to narrow the technological gap. After imposing export controls on Anthropic’s most advanced models and encouraging OpenAI to limit the release of GPT-5.6, the administration reversed course on June 30, 2026, lifting restrictions on Anthropic following weeks of negotiations.

Meanwhile, Chinese companies including 360 Security Technology and Z.ai unveiled new AI models they claim rival leading U.S. systems at lower cost, with some researchers suggesting Z.ai’s capabilities may have been aided by unauthorized distillation of American models. Experts say the episode highlights the challenge of balancing national security with innovation as both countries compete to define the future of AI-powered cybersecurity. → Read the full article here.

COMPANIONS

UBTech Launches Lifelike Humanoid Robot for Home Companionship

Chinese robotics company UBTech has unveiled the U1, a humanoid robot designed for personal companionship rather than industrial work, marking a push into the consumer robotics market. Introduced in Shenzhen on July 1, 2026, the robot features lifelike silicone skin, 88 servo joints, locally running emotional AI powered by Rockchip's RK3588 processor, and on-device data storage for enhanced privacy.

Available in male and female versions, the U1 is offered in Lite, Pro, and Ultra models priced from 119,800 yuan (US$17,650) to 990,000 yuan. The launch reflects a broader effort by Chinese technology companies to bring AI-powered humanoid robots into homes as demand expands beyond factories and warehouses. → Read the full article here.

NEWS

What Else is Happening

Meta Eyes AI Cloud Revenue: Meta will sell excess AI compute, creating a new cloud business to monetize unused capacity and ease spending concerns.

Square Adds AI Restaurant Ordering: Square lets ChatGPT and Claude users order directly from restaurants, avoiding marketplace commission fees.

Portugal Open-Sources National AI Model: Portugal releases Amália, its first national AI model, as fully open source rather than a commercial chatbot.

Together AI Raises $800M: Together AI raised an $800M Series C and secured 500 MW of compute to expand its open-source AI platform.

Venice AI Hits $1B Valuation: Venice AI raised $65M at a $1B valuation after reaching 3M users and a $70M annual revenue run rate.

Want to write for Forward Future?

We're opening up a few contributor slots and want to hear from you. If you've got a take on AI, tech, or where things are headed, we want to know.

Apply Here

POLL RESULTS: REAL OR AI

Can You Still Tell What’s Real?

Yesterday's image was REAL. Here's how you voted: The vote was nearly split, 51% identified it as AI, while 49% believed it was real.

🟨🟨🟨🟨🟨🟨 Real (49%)

🟩🟩🟩🟩🟩🟩 AI (51%)

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

Microsoft Slumps, OpenAI Cuts Costs, & Claude Performance

Matthew Berman — Wed, 01 Jul 2026 11:01:00 +0000

Good morning. It's Wednesday, July 1, and we're covering Microsoft's AI market stumble, OpenAI's major efficiency gains, and Anthropic's push to bring near-frontier AI to more users with Claude Sonnet 5.

Plus: yesterday's poll results on sharing AI-generated wealth, and a guest post on using AI to eliminate permitting bottlenecks in housing.

YOUR DAILY ROLLUP

The ONLY ChatGPT App That Won't Hallucinate

That paper ChatGPT just cited? It might not exist. SciSpace runs real literature reviews inside ChatGPT — every answer backed by verified, clickable sources. It beat Consensus and Elicit across 200 queries on the most relevant papers.

Add SciSpace to ChatGPT FREE — start in 30 seconds!

JNFF40: Get 40% discount on annual plans (limited-time offer)

VIDEO

Getting Started With Hermes AI

Learn how to install and configure Hermes in under 2 minutes, add skills, automate tasks, enable memory, connect Telegram, and more.

FORWARD FUTURE ORIGINAL

AI Won't Build Houses. But It Can Get Them Approved Faster.

You submit your construction plans to the city and wait. Weeks pass, then months. Nobody calls. When you finally hear back, it’s not an approval – you missed one attachment, so you’re back to square one and the bottom of the queue. Too many people know this story, and it’s a big reason housing has become one of the top issues on voters’ minds.

How do we make owning a home more affordable? The conversation usually turns to zoning, funding, density, and land use. Those issues matter. But the bottleneck nobody’s naming sits underneath all of them – the review process itself. And it turns out that’s exactly the kind of problem AI is built to solve. → Read the full article here.

STOCKS

Microsoft Becomes the Biggest Laggard in the AI Stock Rally

Microsoft has emerged as the weakest-performing member of the "Magnificent Seven" in 2026, with its stock on pace to fall 18% in June—its steepest monthly decline since 2000—and down 24% year to date, wiping out roughly $857 billion in market value. Investors are weighing two simultaneous concerns: the company's heavy AI infrastructure spending and growing uncertainty over how AI could disrupt its core software business.

The selloff has pushed Microsoft's forward price-to-earnings ratio to about 21, its lowest level in roughly three years, prompting debate over whether the stock has become undervalued. Investor Michael Burry added to that discussion after disclosing call options that would profit if Microsoft shares climb into the low $700s by 2028, helping spark a 6% rally. → Read the full article here.

EFFICIENCY

OpenAI Finds Inference Optimizations That More Than Halve AI Costs

OpenAI engineers have developed new inference optimizations that more than halve the cost of running existing AI models, according to The Information. The techniques reportedly reduced the number of NVIDIA GPUs needed to serve ChatGPT for signed-out users to just a few hundred, though the company has not disclosed the methods or confirmed the results.

The improvements could allow OpenAI to lower API prices, increase ChatGPT usage limits, or improve profit margins as it continues investing heavily in AI infrastructure. The report highlights that software optimizations—not just adding more chips—are becoming a critical competitive advantage as leading AI companies face ongoing constraints in expanding data center capacity. → Read the full article here. (Paywall)

NEWS

What Else is Happening

Etched Exits Stealth: AI chip startup raised $800M, secured $1B+ in contracts, and plans to ship its first inference racks this summer.

Claude Science Launches: Anthropic's AI workbench unifies research tools, automates workflows, and creates auditable results for scientists.

OpenAI Unveils GeneBench-Pro: New benchmark tests whether AI agents can make complex scientific judgments in computational biology.

Google Launches Nano Banana 2 Lite: Faster, lower-cost AI image generator creates images in four seconds for high-volume workflows.

8090 Labs Raises $135M: Chamath Palihapitiya's AI coding startup secured a Series A led by Salesforce Ventures as he steps in as CEO.

OpenAI Teases Codex Micro Keyboard: OpenAI previewed Codex Micro, a Work Louder shortcut keyboard built for Codex workflows.

REAL OR AI

Can You Still Tell What’s Real?

POLL RESULTS

How Should AI Wealth Be Shared?

Here's how you voted: Opinion was split, with 34% backing a stock dilution model and 32% rejecting the idea of a universal AI dividend entirely.

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

AI Won't Build Houses. But It Can Get Them Approved Faster.

Cyrus Symoom — Tue, 30 Jun 2026 23:54:35 +0000

_{Cyrus Symoom, CEO at Clariti, a leading provider of community development solutions that help local governments modernize permitting operations, simplify staff workflows, strengthen compliance, and deliver faster, more consistent permit outcomes.}

Too many people know this story, and it’s a big reason housing has become one of the top issues on voters’ minds.

Why I Came to Government

I never expected to be writing about permitting. I helped launch Uber Eats across Canada and Latin America and watched technology reshape an industry practically overnight. When I moved into government technology, people asked why. The answer was simple: of every industry I’d seen, government was the furthest behind on technology adoption. That also meant it had the most to gain.

What I found when I got started wasn’t what I expected. Public servants aren’t the problem. Serving a community is genuinely noble work, and most people in government want nothing more than to do it well. Too often, they simply haven’t been given the tools to succeed.

That gap matters more than people realize. For the last two decades, government technology followed one premise: digitize the existing process and call it progress. Move the paper form online. Scan the document instead of filing it. But a broken digital process is still a broken process.

AI creates a genuinely different opportunity: to rethink how work gets done, not just how it gets processed. We often talk about AI as a tool for writing code or answering questions. What gets less attention is its ability to streamline complex administrative systems that handle huge amounts of information. Permit reviews are a perfect example: hundreds of pages of plans checked against codes, prior submissions, and dozens of rules at once – high-volume work that has overwhelmed reviewers for decades, and exactly what AI can fix.

But that potential only materializes when we hold these systems to a higher bar: real outcomes, not just a process moved online. If review times aren’t measurably improving, the technology isn’t working.

This also means being clear-eyed about what AI can and can’t do. It isn’t a cure-all. It won’t fix zoning. It won’t create more funding for housing. But it can fundamentally improve the review layer, catching incomplete applications before they enter the queue, reducing the correction cycles that send projects back to the end of the line, and freeing staff to focus on decisions instead of paperwork.

What This Looks Like in Practice

If this sounds theoretical, look at what’s happening in Honolulu. City leaders have set an ambitious goal: build one of the fastest permitting processes in the country. Residential plan review times have fallen 70%, and applicants who previously waited 73 days for a response now hear back in 32.5 days. That adds up when every application contains hundreds of pages of plans, forms, and supporting documents that have to be reviewed for completeness before anyone can make a decision. The city’s planning director described the approach as “TurboTax for permitting” – applicants are guided toward complete, compliant submissions before a reviewer ever opens the project.

But the technology alone doesn’t explain those results. A lot of the progress came from rethinking how the review process worked in the first place. That's where AI is different. Earlier tools just digitized the old process. AI is the first to spark conversations about reimaging them entirely.

Those improvements have real material consequences. When projects sit in review for months instead of days, costs rise, investors lose confidence, and housing that communities desperately need never gets built. Faster, more predictable reviews help projects move forward, businesses open their doors, and families get permit decisions sooner.

Why This Matters Now

The political moment has caught up with public frustration. A White House executive order, bipartisan legislation just passed in Congress, governor-led initiatives across the country, and a new $3 million federal grant for automated permitting systems all point in the same direction: permitting reform is becoming a national priority.

We’re seeing the shift firsthand. Cities are coming to us trying to figure out where AI fits in their operations – and increasingly it’s not just the big metros, but smaller cities that have watched backlogs grow without the staff to clear them.

But none of that pours a single foundation or frames a single house. Funding and mandates only go so far. The cities making the most progress aren't just buying new software, they're rethinking how the work gets done. Once online banking became mainstream, waiting in line felt unacceptable. Once on-demand services became popular, slow and unpredictable became inexcusable. AI is creating that same shift in expectations for government.

This isn’t really a story about permitting. It’s about what becomes possible when public servants finally have tools that match the complexity of the work they’re doing. It’s about everyone waiting on the other side of the process: the family that finally breaks ground, the business that opens on time, the homes a community has needed for years being built. The cities moving first won’t just approve permits faster. They’ll show what it means to reinvent government, not just digitize it.

Cyrus Symoom is CEO of Clariti, a government technology company helping local governments modernize permitting, licensing, and community development services. Since acquiring the company in 2020, Cyrus has helped grow Clariti from a small software business into a platform used by hundreds of government agencies across North America.

Prior to Clariti, Cyrus led Uber Eats’ expansion across Canada and Latin America, launching the business in multiple countries and cities throughout the region. He began his career in investment banking at CIBC Capital Markets before co-founding investment firm Explora Partners.

Today, Cyrus works with cities across North America on housing, permitting reform, AI adoption, and government service delivery.

Amazon-Anthropic Deal, Meta Delays, & AI Cooling

Matthew Berman — Tue, 30 Jun 2026 11:01:00 +0000

2026 State of AI in the Enterprise: Box’s Report Available Now!

Good morning. It's Tuesday, June 30, and we're covering growing tensions in the Amazon–Anthropic partnership, why AI data centers are investing in smarter cooling, and how compute shortages are slowing Meta's AI.

Plus: a guest essay arguing for a new way to share the wealth created by automation, and a reader poll on the same topic.

YOUR DAILY ROLLUP

AI Leaders Are Betting on Platform Flexibility

Most organizations are concerned about AI provider lock-in and know it’s critical that their agents operate “headlessly” — connecting directly to systems, APIs, and data sources without a human interface.

Box’s State of AI report breaks down how the companies on the leading edge of AI are betting on infrastructure flexibility rather than a winning model.

Download the report to learn more!

VIDEO

Regulation Reshapes Frontier Access

FORWARD FUTURE ORIGINAL

A Tax System Built on Labor Is a Tax System Built on a Melting Glacier

Something strange happened over the past ten months. The idea that the public should own a stake in the companies getting rich off automation went from fringe to seeming consensus, fast.

POLL

How Should AI Wealth Be Shared?

Several "universal AI dividend" plans are gaining traction. Sanders' 50% stock tax, Trump's sovereign wealth fund equity grabs, Altman's old share-tax plan.

Scott Santens proposes something different: a never-ending 1.5% stock "dilution tax" that mints and sells new shares every year, paying ~$365/month to every American, with no government ownership stake.

PARTNERSHIPS

Amazon Renegotiates Anthropic Deal as AI Costs Rise

Anthropic has renegotiated part of its partnership with Amazon, shifting pricing for its AI models from compute-hour billing to token-based billing starting in 2027, according to people familiar with the discussions. The change could increase Amazon's costs for products powered by Claude, prompting the company to evaluate alternatives, including OpenAI's models and its own Nova family, although Amazon disputes that the revised agreement will raise its overall costs.

The development highlights growing tension in one of AI's most significant partnerships as Amazon balances major investments in both Anthropic and OpenAI while seeking greater control over the technology underpinning its AI services. → Read the full article here. (Paywall)

FUNDING

Omen AI Raises $31 Million to Monitor Data Center Cooling Systems

Omen AI has raised a $31 million Series A funding round led by Nava Ventures to expand its real-time fluid monitoring technology for AI data centers. The startup's miniature spectrometer detects bacterial growth and other contaminants in liquid cooling systems before they cause clogs and costly downtime, which can sideline GPU racks for several hours.

Originally built for heavy machinery maintenance, the technology gained traction through Caterpillar dealerships before the company pivoted toward the fast-growing AI infrastructure market. → Read the full article here.

NEWS

What Else is Happening

Proception Settles Tesla Suit: Robot hand startup settled Tesla's trade secret lawsuit, raised $11M, and began shipping its first robotic hands.

Ford Rehires Quality Engineers: Ford brought back 300+ veteran inspectors after AI quality checks fell short, highlighting the limits of automation.

Google Opens Gemini Images to Free Users: Free US users can now generate personalized images via Gemini's Nano Banana model.

Momenta Seeks $751M IPO: GM- and Tencent-backed self-driving startup launched a $751M Hong Kong IPO to fund AI R&D and Robotaxi expansion.

Pocket Raises $11M: AI note-taking startup Pocket raised $11M after selling 130,000 recording devices for meetings and transcription.

TIDAL Demonetizes AI Music: TIDAL will stop paying royalties on fully AI-generated tracks and remove AI impersonations to prioritize human artists.

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

A Tax System Built on Labor Is a Tax System Built on a Melting Glacier

Scott Santens — Mon, 29 Jun 2026 23:49:12 +0000

_{Guest contributor: Scott Santens writes Foreword for Humanity on Substack, advocates for Universal Basic Income, and founded the Income To Support All Foundation and the AI Pledge for Humanity.}

Something strange happened over the past ten months. The idea that the public should own a stake in the companies getting rich off automation went from fringe to seeming consensus, fast.

Bernie Sanders introduced the American AI Sovereign Wealth Fund Act, which would hit the biggest AI companies with a one-time 50% tax paid in stock and drop those shares into a fund that pays dividends to all of us. Trump signed an executive order creating a sovereign wealth fund, and his administration has been grabbing equity stakes along the way: in Intel, in MP Materials, in US Steel, and more recently with the big AI companies, with a household dividend floated as one use of the proceeds. Venture capitalist Vinod Khosla floated taking 10% of every public company for the people. And back in 2021, Sam Altman proposed almost exactly this in "Moore's Law for Everything": an American Equity Fund capitalized by taxing big companies 2.5% of their market value each year, payable in shares.

So the question is no longer whether the public should own a piece of the wealth we all helped build. People across the political spectrum have conceded that point. The question now is how.

And on the how, I think nearly everyone is making the same mistake.

They all want to build a giant pile of wealth.

I don't. I propose we build a faucet.

How To Build a Different Kind of Wealth Fund

Every one of the aforementioned proposals ends the same way: a huge pile of stock sitting in a government fund, like Norway's or Alaska's, growing over time while possibly paying out a thin slice of itself each year. That's a fine model if your goal is to build a mountain of wealth.

But is that the best way to do it?

A fund that holds shares makes the government a permanent shareholder — picking which companies get a federal partner, accumulating votes and board seats, becoming both referee and player in markets it also regulates. We're watching that happen in real time, with equity grabbed selectively from some companies and not others. The machinery one administration builds to do good, the next one inherits to reward friends and punish enemies.

So here's my alternative. Require every large company above $100 million in revenue or value (public or private) to issue new shares equal to 1.5% of its value each year into a public fund. Not a check. Not the sale of a factory. Just newly minted stock, the same way companies already mint shares to pay executives or swallow rivals. At today's valuations that's about $1.76 trillion a year, or roughly $365 a month for every American, adult and child alike.

Then the fund slowly and steadily sells everything it receives over the next year. There are about 250 trading days in a year, so each day it sells 1/250th, dripped slowly across the day the way an algorithm quietly works a large order. The goal is to be the most boring participant in the market. The next year, the companies refill it, and we do it again.

Won't daily selling crater the market? We already run this exact machine in reverse. American companies spent about $1.1 trillion buying back their own stock in 2025, an all-time record, pulling shares out of the market to enrich the people who already own them. My dilution tax proposal is simply buybacks run backward: selling shares to enrich everyone instead. If the market can swallow a trillion dollars of buying, it can absorb a comparable amount of selling, especially as a slow, predictable daily drip.

And because the fund sells everything, it never builds a lasting stake in anyone. The public owns the wealth without the government owning the control. A wealth fund that hoards nothing and controls nothing — a wealth fund that never becomes a power grab.

It's also the most inescapable wealth tax ever designed. When you mint 1.5% more shares, every existing share is worth a hair less. It's inflation, but for stocks. Move to an island, hire the cleverest lawyers, route everything through shell companies — none of it matters. You cannot flee dilution. You cannot deduct it. It finds you wherever you are. And it's gentle: at 1.5%, the very wealthy don't lose wealth, they just compound a little slower — maybe 8.5% a year instead of 10%. They still get richer. They just share a sliver on the way up.

Now Tax the Transactions

Our whole tax system leans on labor. Payroll and income taxes supply the bulk of federal revenue. So what happens as AI and robots do a growing share of the work? When a machine takes a job or just could, the paycheck disappears or shrinks and so does every tax that paycheck carried. The glacier is melting, and our entire tax system is standing on it.

The fix is to tax the things automation can't shrink and will instead grow: wealth and transactions. The stock dilution tax handles the wealth. A consumption tax handles the rest.

A value-added tax is a consumption tax already operating in every OECD nation except the US, but the variant I find most interesting is the Automated Payment Transaction tax as proposed by economist Edgar Feige — a tiny levy on every payment that clears the banking system. Every purchase, every trade, every wire, collected automatically by software at the moment of settlement. No forms, no returns. Fedwire alone moves over $1 quadrillion a year; add the card networks and the rest and US payment volume tops $1.5 quadrillion which is more than eighty times the consumption a sales tax would reach. A rate of roughly 0.12% on each side of a transaction raises as much as a 10% national VAT. Buy a $5 coffee, pay half a cent. Execute a $50 million block trade, pay $60,000 — which is exactly the point, because most transaction volume is finance, not coffee. The base is more skewed toward the rich than income itself.

And this already exists to a degree. Every time you swipe a card, the networks already skim 2% to 3% — automatically, invisibly, in code. The entire payments system is already a transaction-tax machine. It just collects for shareholders instead of for the public.

AI and robots don't earn wages but they are bought, and they do transact. A transaction tax captures the torrent of machine-to-machine payments the agentic economy is already starting to generate: software autonomously buying compute, data, and services from other software, billions of times a day, at speeds no payroll tax will ever touch. Human labor may decline as a share of the economy. Transactions won't. They'll explode. We should tax the thing that grows.

Multiple Paths, One Floor

Neither of these is the single right answer, and that's the point. They stack.

The stock faucet would pay about $365 a month. A transaction-tax dividend can pay roughly $485 more. Add a refundable $375/mo tax credit replacing the standard deduction — another layer I've written about — and you're past $1,200 a month for every citizen and green card holder, adult and child, before anyone earns a dollar of wages. Hike the stock dilution tax and the transaction tax a bit and you reach a number that more than ends poverty in America and significantly reduces income and wealth inequality.

The people building AI keep telling us something like a universal dividend will be necessary. Elon Musk calls it "universal high income." Google DeepMind's Demis Hassabis says much the same. Fine. Let's hold them to it and let's fund it the right way, by taxing what AI grows instead of clinging to the personal income tax base it's shrinking.

We trained these machines on everything any of us ever wrote, living and long dead. We funded the science they stand on, going back generations. We are the data, the customers, and the reason any of it has any value at all. A dividend isn't charity, and it isn't radical. It's ours.

So let's not build another giant pile of stock. Let's build the faucet — and finally turn it on.

Scott Santens writes Foreword for Humanity on Substack and advocates for Universal Basic Income. He is the founder and CEO of the Income To Support All Foundation and the creator of the AI Pledge for Humanity, which calls on AI leaders to back their words about UBI with action.

AI Job Fears, Custom AI Chips, & GPT-5.6 Government Delay

Matthew Berman — Mon, 29 Jun 2026 11:01:00 +0000

Good morning. It's Monday, June 29, and we're covering why Americans still don't trust AI, why Anthropic is fielding questions over model security, and why OpenAI is rethinking its IPO.

Plus: a prompt that turns AI chatbots into your personal tutor, and a guest essay exploring why there’s no going back to the pre-ChatGPT era.

YOUR DAILY ROLLUP

Claude Tag's Hidden Strategy

Anthropic's Claude Tag turns Slack into an AI coworker, raising concerns over AI-native companies, vendor lock-in, and the future of knowledge work.

FORWARD FUTURE ORIGINAL

Genie’s Not Going Back in the Bottle

I was recently watching a Demis Hassabis interview, and his one line has stuck with me. When Stanford President John Levin asked, “Demis, if you were back in school, what would you be studying, and what would be your advice to students about what to study in their careers?”

I want to stop here for a moment. Because this question relates to a specific pattern that has emerged among university students in the last month, and it shows what’s on their minds regarding what would be the impact of AI on their professional lives. → Read the full article here.

JOB SECURITY

Americans’ AI Gloom Reflects Job Security Fears, Not Messaging

Americans are more pessimistic about artificial intelligence than people in nearly every other country, according to a survey of 24,000 adults across 30 nations. Venture capitalist Paul Kedrosky argues this skepticism is driven less by misinformation or cultural attitudes than by concerns over job security in advanced economies.

In countries with large informal workforces, AI is often viewed as a path to better economic opportunities, while in wealthier nations with stable labor markets, it is seen as a threat to employment, income, and professional status. The piece suggests that Americans’ unease stems from the risk of losing what they already have, rather than resistance to technology itself. → Read the full article here. (Paywall)

MEMORY BOOM

Wall Street Crowns Micron the Next NVIDIA

Micron briefly leapfrogged Meta and Tesla in market value Thursday, with shares up more than 236% in a single month after blockbuster earnings: revenue quadrupled to $41.45 billion and profits jumped from $1.88 billion to $28.2 billion year-over-year. The Idaho memory maker is riding an AI data-center buildout that has created a severe shortage of DRAM, NAND, and especially High-Bandwidth Memory, the kind AI servers consume in massive quantities.

To calm fears of the boom-and-bust cycles that have historically plagued memory makers, Micron is pointing to 16 long-term supply agreements with customers including NVIDIA and Anthropic. Analysts are buying the story, though whether Micron can dodge the next downturn remains the open question hanging over its trillion-dollar run. → Read the full article here.

ARCHITECTURE

Former Databricks AI Chief Unveils 1,000x Lower-Power AI Architecture

Unconventional AI, a startup led by former Databricks AI chief Naveen Rao, has introduced Un-0, an image-generation model that runs on a software simulation of the company's new oscillator-based computing architecture. The system matches the performance of conventional diffusion models while serving as an early proof of concept for hardware Rao says could reduce AI inference power consumption by up to 1,000 times.

The company plans to release chip schematics and eventually build a complete inference platform around the technology. While the energy-saving claims remain unproven on physical hardware, the effort reflects growing industry interest in rethinking AI infrastructure as power demand becomes a key constraint on scaling. → Read the full article here.

NEWS

What Else is Happening

Mark Cuban Urges AI Outreach: Cuban says AI firms must support workers, artists and communities or opposition to data centers will keep growing.

OpenAI Names India MD: Former Uber India president Prabhjeet Singh will lead OpenAI's India expansion, its biggest market outside the U.S.

Ukraine Plans AI Infrastructure: Kyivstar will help build domestic AI computing to support military needs while reducing reliance on foreign infrastructure.

DuckDuckGo AI Spreads Hoax: DuckDuckGo's AI falsely claimed Trump died of rabies after repeating fabricated online content, exposing AI misinformation.

PROMPT OF THE WEEK

The "Test Me" Prompt

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

AI Energy Grids, Google's Coding Push, & AI Model Theft

Matthew Berman — Fri, 26 Jun 2026 11:01:00 +0000

Good morning. It's Friday, June 26, and we're covering how homes could help power AI data centers, Google's race to catch Anthropic in AI coding, and escalating allegations over Alibaba’s AI model distillation.

Plus: yesterday's poll results on China's robotics lead, and a Friday Fact about something with 10× more DNA than humans.

YOUR DAILY ROLLUP

Their DNA Dwarfs Yours by 10x

A creature small enough to fit in a cereal bowl is carrying around a genetic library ten times the size of a human's. Scroll to the bottom to find out which one.

VIDEO

Codex Clones Excel From Scratch

A six-word prompt had Codex autonomously clone Microsoft Excel over 12 days, recreating formulas, formatting, sorting, and core features.

ENERGY

Tesla, Sunrun, and Renew Home Aim to Power AI Data Centers With Homes

Tesla, Sunrun, and Renew Home announced on June 24, 2026, that they are partnering to use rooftop solar panels, home batteries, smart thermostats, and other connected devices to help meet the soaring electricity demands of AI data centers. By coordinating millions of residential energy devices into virtual power plants, the companies say they can free up enough grid capacity to support the equivalent of 17 large data centers during peak demand, potentially avoiding years of new infrastructure construction.

The initiative would also compensate participating households through bill credits or cash payments while reducing pressure on overstretched power grids. Success will depend on utilities, grid operators, and regulators embracing the model beyond existing programs in states like California. → Read the full article here. (Paywall)

COMPETITION

Google Reorganizes AI Coding Team as Anthropic Pressure Mounts

Google is restructuring its AI coding strike team into a permanent "midtraining" organization as it works to close the gap with Anthropic in AI coding, one of the industry's fastest-growing and most profitable markets. The reorganization comes amid the departures of prominent AI researchers Noam Shazeer, who joined OpenAI, and Nobel laureate John Jumper, who moved to Anthropic, raising fresh questions about Google's talent retention and compute allocation strategy.

The new team will focus on improving model capabilities through an intermediate training stage between pretraining and post-training, with the goal of accelerating performance in coding and other enterprise tasks. The changes highlight Google's effort to remain competitive as rivals rapidly expand AI tools beyond software development into broader workplace applications. → Read the full article here. (Paywall)

SEMICONDUCTORS

IBM Unveils Sub-1 Nanometer Chip Technology for AI Era

IBM on June 25, 2026, unveiled what it says is the world's first chip technology capable of producing transistors smaller than one nanometer, introducing a 0.7-nanometer "nanostack" architecture designed to boost AI computing performance. The company says the new design fits nearly 100 billion transistors onto a fingernail-sized chip, delivering up to 50% higher performance or 70% greater energy efficiency than its 2-nanometer technology announced in 2021.

IBM also said the architecture reduces SRAM memory size by 40%, an advance that could benefit AI accelerators that rely heavily on on-chip memory. The technology is expected to reach production within five years, though IBM has not yet named a manufacturing partner, as competition intensifies among IBM, Intel, TSMC, and other chipmakers to meet growing AI demand. → Read the full article here.

NEWS

What Else is Happening

Qualcomm Raises AI Revenue Target: Shares rose after Qualcomm doubled its revenue goal to $40B by 2029, including $15B from data centers.

Apple Raises Mac, iPad Prices (Paywall): Apple increased some prices by up to $300, citing higher memory and storage chip costs driven by AI demand.

Micron Warns of Memory Shortages: Micron expects demand to keep DRAM and NAND supplies tight beyond 2027, sustaining higher memory prices.

Apple Shifts Mac Chip Strategy (Paywall): Apple plans to skip high-end M6 Mac chips, instead focusing its next processors on an AI-focused M7 generation.

Amazon Expands India AI Investment: Amazon will invest another $13B in India's AI and cloud infrastructure, raising its 2030 commitment to $48B.

FRIDAY FACT

The Axolotl's Genome Is 10 Times Larger Than Ours

The axolotl, a Mexican salamander known for regrowing limbs, organs, and even parts of its brain, carries about 32 billion base pairs of DNA. The human genome, by comparison, comes in around 3.1 billion. That makes the axolotl's genome the largest ever fully sequenced, roughly ten times our size by raw DNA length.

The size gap isn't about complexity. Axolotls have around 23,000 protein-coding genes, close to the human count of about 20,000. The bulk of their extra DNA is repetitive sequences, including a huge expansion of transposable elements that don't code for proteins at all. Researchers are now combing through that bloated genome looking for clues to how the axolotl regenerates tissue, a trick humans can't pull off, no matter how much DNA we're packing.

POLL RESULTS

Will China Lead the Future of Robots?

Here's how you voted: 86% believe China is the clear winner in the robotics race, while 14% think the U.S. or another challenger could still take the lead.

🟩🟩🟩🟩🟩🟩 Yes — China is the clear winner (86%)
⬜️⬜️⬜️⬜️⬜️⬜️ No — The US will win eventually (8%)
⬜️⬜️⬜️⬜️⬜️⬜️ Other – A new player will take the crown (6%)

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

Genie’s Not Going Back in the Bottle

Deepanshu Sharma — Fri, 26 Jun 2026 00:15:58 +0000

_{Guest contributor: Deepanshu Sharma is an independent analyst, researcher, and writer decoding AI capabilities, evaluating benchmarks, and simplifying the societal impacts of these emerging technologies.}

I was recently watching a Demis Hassabis interview, and his one line has stuck with me. When Stanford President John Levin asked, “Demis, if you were back in school, what would you be studying, and what would be your advice to students about what to study in their careers?”

I want to stop here for a moment.

Because this question relates to a specific pattern that has emerged among university students in the last month, and it shows what’s on their minds regarding what would be the impact of AI on their professional lives.

Many prestigious people received backlash on stage from students while giving commencement speeches for discussing and utilizing artificial intelligence. It reflects a growing perception among students that AI will wipe out all Jobs.

Current evidence states otherwise. For example, let’s take GDPval, which is an evaluation framework introduced by OpenAI in September 2025, that covers 44 occupations across the top 9 industries. On the one hand, it evaluates the latest AI models on real-world tasks, but on the other hand, its results also tell us why a human is required in the loop.

GDPval measures the performance of an LLM on 1320 specialized tasks. It is presented as a benchmark that truly represents the capabilities attained by LLMs, comparing them directly to the everyday work people actually do.

In the original paper, Claude Opus 4.1 acheived 47.6%, which was shown as almost reaching the level of an industry expert. On the surface, it looks like LLMs are good enough to compete with experts and gain general intelligence as compared to humans, who are restricted to expertise only in their own domain.

Initial Assessment of Frontier Models from the Original Paper

Reality is slightly twisted, and there are reasons for it.

First is the limited scope of work that can actually be done by LLMs. This benchmark evaluates the performance of LLMs solely on “well-specified digital tasks”. If I were to summarize it:

This benchmark does not cover all the tasks of any single occupation. Only a part of them.
Then, from that subset, it only covers tasks that can be done digitally
Even then, it only includes tasks that can be clearly specified from the start, meaning it cannot cover any scenario where information changes midway through.

Limitations from the Original GDPval Paper.

Secondly, I want to emphasize the speed and cost improvements.

If these systems can complete a task faster than a human, then it makes sense to use them. However, it completely defeats the purpose if reviewing their work ends up taking more time overall.

According to the initial paper, only GPT-5 shows some meaningful speed improvements. In fact, looking at the data, using GPT-4o actually slowed the work down.

Furthermore, this evaluation ignores the impact of catastrophic mistakes, instances where the cost of a single error can wipe out any savings gained from using AI in the first place. The paper highlights examples such as insulting a customer, giving a wrong medical diagnosis, recommending fraud, or suggesting actions that could cause physical harm. According to their own evaluation, this occurs 2.7% of the time.

Speed & Cost Improvement Table from the Original GDPval Paper.

Two weeks ago, Sam Altman clarified on this when an interviewer asked, “When you introduced GPT-5.2, you said it outperforms professionals across 44 occupations. You can understand why there may be a backlash when people hear things like that.”

Sam said:

“What I wish we had said then is that it outperforms professionals at small tasks in 44 occupations, which is, I think, a more accurate thing”

Excerpts of GDPval from GPT-5.2 release notes.

Discussion on GDPval and other stats (like GPT-5.2) was very important because we still use this benchmark to test the latest models, and this was one of the reasons for the prevailing misconception that AI will soon wipe out all jobs. People rarely remember the actual benchmark; instead, the emotional narrative spreads that AI is already beating experts in every domain.

I wanted to clear the air. But in AI, one month feels like one year.

So much has changed in just the last couple of days, and I want to shift focus to the latest models like Fable 5 and how they handle real-world complexities.

On June 19th, Artificial Analysis introduced a new benchmark called AA-Briefcase. It tests models on agentic knowledge work involving practical complexities, such as multi-week projects requiring thousands of inputs (company documents, meeting transcripts, large-scale data exports, Slack messages, and emails).

Fable 5 performed better than other models; it successfully completed only 3% of the tasks. Furthermore, on 31 out of 91 tasks, not a single model scored above 50% on the rubric criteria.

This shows the capabilities of these models in real-world settings. If we focus on why these scores are low, the reasons were:

Strong models fail by missing nuanced requirements hidden within the task or source files.
Less capable models struggle with basic execution, often ignoring input files, delivering unusable work, or producing nothing at all.

It’s true that AI is far from eliminating professionals. However, every profession has so many sub-tasks where AI significantly outperforms us, and failing to automate those will be a costly mistake.

I know AI can summarize multiple academic papers faster than I can, extract data from long documents, find insights I might otherwise miss, and help me understand new perspectives.

The list goes on.

AI is the exact reason I am excited about learning new things. Now, I don’t have to worry about not knowing the basics of a subject. Research is only one prompt away.

In my opinion, two types of people will win with AI. The first are those who are already experts in their domain (or intend to become experts); they know exactly which tasks to delegate to AI, freeing them up to work on the rest and innovate. The second are those who will use AI to rapidly learn new things, specifically focusing on vertical integration to expand their footprint into new domains.

I want to go back to Demis’s answer to John, and I’ve never heard a better explanation on this sensitive subject from anyone else, Demis said:

He further says, “I would also lean in, though, to not wish it away. The genie's not going back in the bottle.”

It’s true that we can’t go back to the pre-AI era, whether in our personal or professional lives. AI will seep in more and more as time passes. The only way to tackle this generational challenge is to get this technology on our side. We must find better ways to execute our next big task, so that it takes substantially less time than before, or completely eliminate the tasks that AI is already good at.

I hope I was able to add a little value to your day. Thank you so much for your time, and have an absolutely wonderful day!

Deepanshu Sharma is an independent analyst, researcher, and writer decoding AI capabilities, evaluating benchmarks, and simplifying the societal impacts of these emerging technologies.

China's Supercomputer, Data Center Backlash, & Gemini's Automation

Matthew Berman — Thu, 25 Jun 2026 11:01:00 +0000

Good morning. It's Thursday, June 25, and we're covering China's GPU-free supercomputing, the U.S. political battle over AI infrastructure, and Google's push to make AI agents more autonomous with Gemini 3.5 Flash.

Plus: the verdict from yesterday's Real or AI challenge.

YOUR DAILY ROLLUP

12 Open-Source AI Tools Worth Trying

Discover 12 standout open-source AI projects, from autonomous agent frameworks and coding copilots to OCR models and voice AI tools.

POLL

Will China Lead the Future of Robots?

Morgan Stanley just doubled its forecast for China's humanoid robot shipments this year to 50,000 units — and projects 446,000/year by 2030.

SUPERCOMPUTING

China Reclaims Supercomputing Lead With GPU-Free LineShine System

China’s LineShine supercomputer, located in Shenzhen, was named the world’s fastest on June 23, 2026, ending the United States’ hold on the top spot and marking China’s first No. 1 ranking since 2017. According to Top500 benchmark tests, LineShine outperformed the U.S.-based El Capitan system by more than 20%, despite relying solely on central processing units (CPUs) rather than graphics processing units (GPUs), the chips that power most leading AI and supercomputing systems.

Experts say the architecture could offer a new approach to combining artificial intelligence workloads with traditional scientific computing, while also reducing dependence on advanced AI chips targeted by U.S. export restrictions. The achievement highlights China’s ability to advance high-performance computing through alternative chip designs and comes amid intensifying technological competition between Washington and Beijing. → Read the full article here. (Paywall)

INFRASTRUCTURE

America’s AI Data Center Boom Faces Growing Local Backlash

A wave of opposition to AI data centers is spreading across the United States as tech companies pour hundreds of billions of dollars into the infrastructure needed to train and run advanced AI models. Residents and local officials are increasingly pushing back over concerns about noise, land use, power consumption, environmental impacts, and the rapid industrialization of communities, leading to the cancellation of at least $42 billion in projects during the first quarter of 2026 alone.

The backlash comes as companies including Meta, Microsoft, Google, OpenAI, and Amazon race to expand computing capacity, with industry forecasts suggesting trillions of dollars will be invested globally in AI data centers by 2030. The growing resistance has turned data centers into a political issue, raising questions about whether the infrastructure required to sustain the AI boom can be built quickly enough to meet demand and maintain America’s technological lead over China. → Read the full article here. (Paywall)

NEWS

What Else is Happening

AI Reads Vesuvius-Burnt Scroll: Researchers used AI to reveal 20 columns of hidden Herculaneum text on Stoic philosophy without unrolling it.

Qualcomm Buys AI Startup Modular: Qualcomm acquired AI startup Modular to strengthen its software stack and expand its data center business.

Cerebras Shares Fall After Earnings: Cerebras shares fell after it forecast lower profit margins than AI chip rivals, overshadowing strong revenue growth.

Agility Robotics Targets $2.5B SPAC Deal: Agility Robotics plans a $2.5B SPAC merger to scale Digit production and meet growing demand for robots.

Anthropic Poaches More Google AI Talent: Two key researchers are leaving Google for Anthropic, adding to a string of high-profile talent departures.

POLL RESULTS: REAL OR AI

Can You Still Tell What’s Real?

Yesterday's image was AI. Here's how you voted: 75% spotted it as AI, while 25% were convinced it was real.

🟨⬜️⬜️⬜️⬜️⬜️ Real (25%)

🟩🟩🟩🟩🟩🟩 AI (75%)

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

Data Center Economics, China's AI Leap, & Meta Arena

Matthew Berman — Wed, 24 Jun 2026 11:01:00 +0000

Good morning. It's Wednesday, June 24, and we're covering the economic ripple effects of AI data centers, China's latest push to close the AI gap with the U.S., and Meta's experiment with prediction markets.

Plus: yesterday's poll results, and a new test of your ability to separate reality from AI-generated content.

YOUR DAILY ROLLUP

AI Loops Optimize App Speed

Self-improving AI continuously benchmarks and optimizes app performance, moving page by page until all load times fall below 50ms.

ECONOMICS

AI Data Centers Bring More Local Benefits Than Critics Assume

Columnist Ann Davis Vaughan argues that many concerns about AI data centers overlook their growing economic impact on host communities. She points to examples including rural Louisiana, where tax revenue linked to a large Meta campus helped fund $50,000 teacher bonuses, and West Memphis, Arkansas, where local officials say Google's investment has boosted business recruitment. While acknowledging concerns over power, water, tax incentives, and housing, Vaughan cites studies projecting thousands of long-term jobs from data center projects in Wisconsin and Texas.

She also notes that regulators are increasingly requiring tech companies to cover infrastructure costs, while communities have gained leverage to negotiate investments in schools, roads, and public services. The piece contends that growing opposition to data centers may cause some communities to miss economic development opportunities, even as developers face rising pressure to prove they can deliver on promised local benefits. → Read the full article here. (Paywall)

AI RACE

China’s Latest AI Model Narrows Gap With U.S. Rivals

Chinese AI startup Zhipu released its new open-source model, GLM 5.2, on June 13, claiming performance close to leading American systems at a fraction of the cost. Benchmark rankings place it among the strongest AI models available, with some evaluations showing it outperforming newer U.S. competitors on workplace-oriented tasks. However, researchers caution that Chinese models often score better on public benchmarks than private tests, where American firms such as OpenAI and Anthropic appear to maintain a lead of roughly eight to ten months.

The release comes as U.S. AI access restrictions and rising usage costs push some companies to explore alternatives, though lower advertised prices do not always translate into lower real-world costs because Chinese models typically consume far more computing resources. The development suggests China is closing the AI gap faster than many expected, even as American companies retain an edge in overall capability and efficiency. → Read the full article here. (Paywall)

ROBOTICS

MIT Chip Enables Real-Time 3D Mapping on 6 Milliwatts

MIT researchers have developed a system-on-a-chip called Gleanmer that allows small robots and other battery-powered devices to build detailed 3D maps in real time while consuming just 6 milliwatts of power—roughly the energy used by a single LED. Instead of relying on memory-intensive voxel-based mapping, the chip uses flexible Gaussian representations and a custom algorithm that processes depth images in a single pass, dramatically reducing memory and energy requirements.

The system consumes about 2.5% of the power required by the best existing mapping chips and can cut navigation energy use by roughly 80% by reusing map data during path planning. The technology could enable tiny drones to inspect infrastructure, improve lightweight augmented reality headsets, and expand the capabilities of autonomous devices operating with strict power constraints. → Read the full article here.

NEWS

What Else is Happening

Google’s AI Era Cracks Show: ChatGPT, Bing, and DuckDuckGo are gaining users as Google's search traffic dips, exposing new cracks in its core business.

Meta Launches $299 AI Glasses: Meta launched its first in-house smart glasses, expanding beyond partnerships with Ray-Ban and Oakley.

Engram Raises $98M for AI Memory: The AI memory startup says its technology can cut token use by up to 100x, helping companies lower AI costs.

SK Hynix Overtakes Samsung: AI-driven demand for memory chips pushed SK Hynix past Samsung in market value after a historic stock rally.

Meta Pauses Employee AI Training: Meta halted an AI training project using employee keystrokes after a leak exposed internal conversations and data.

REAL OR AI

Can You Still Tell What’s Real?

POLL RESULTS

Can We Afford to Slow AI Down?

Here’s how you voted on yesterday’s poll: 38% believe AI’s pace is beyond our control, while 25% say the right approach depends on the domain.

🟨🟨🟨⬜️⬜️⬜️ Speed — the benefits outweigh the risks (22%)
🟨🟨⬜️⬜️⬜️⬜️ Caution — slow down, even if it costs us (14%)
🟨🟨🟨🟨⬜️⬜️ Depends on the domain (25%)
🟩🟩🟩🟩🟩🟩 We can't actually choose the pace either way (38%)

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

Chinese AI Rise, NVIDIA's Cooling Innovation, & SpaceX's AI Revenue

Matthew Berman — Tue, 23 Jun 2026 11:01:00 +0000

Good morning. It's Tuesday, June 23, and we're covering the rapid rise of Chinese AI models, NVIDIA’s effort to slash data center water consumption, and how SpaceX is turning AI compute demand into a major business.

Plus: a poll on AI acceleration, and a guest essay on why building the loop matters more than building the agent.

YOUR DAILY ROLLUP

Building a DIY Bitzee With AI

A custom Bitzee powered by Raspberry Pi and Codex transforms a $40 virtual pet into a programmable touchscreen companion.

GUEST POST

Build the Loop, Not the Agent

^{_{Guest contributors: Ethan Wang is founding member of Google DeepMind Spark agent and founding member of Google DeepMind Mariner Agent. Namrata Ganatra is VP of Product & Engineering at Intuit, founder of AI-Native Builders, former AI startup founder, and ex-leader at Meta, Coinbase, Xero, and Microsoft.}}

Most teams treat agent modernization as a project with a finish line: re-architect around today's frontier model, ship, declare victory. But model capability advances faster than any single modernization effort can complete. By the time you finish re-architecting around today's frontier model, the next one has already shifted the ground beneath you.

Plan around a fixed end state and you will perpetually ship an architecture tuned to a model generation that is already obsolete. The better bet is simple: the team that can iterate fastest against the newest models will win. → Read the full article here.

AI RACE

Chinese AI Models Gain Ground as U.S. Firms Embrace Lower-Cost Alternatives

Chinese artificial intelligence developers are narrowing the performance gap with leading U.S. labs faster than many industry observers expected, driven by a wave of powerful open-source models from companies including DeepSeek, Moonshot AI, and Z.ai. Microsoft is reportedly considering offering DeepSeek’s V4 model through its Copilot ecosystem as a lower-cost option to models from OpenAI and Anthropic, while developers increasingly adopt Chinese models through platforms such as OpenRouter.

Recent releases including Moonshot’s Kimi and Z.ai’s GLM-5.2 have drawn praise from Silicon Valley executives, with some claiming performance comparable to top American systems. The trend highlights a growing tension between Washington’s efforts to limit China’s AI advances and the willingness of businesses to prioritize cost, accessibility, and performance when choosing AI tools. → Read the full article here. (Paywall)

SUSTAINABILITY

NVIDIA Unveils Data Center Cooling That Eliminates Water Consumption

NVIDIA says its new liquid-cooling architecture can reduce operational water use in AI data centers by up to 100% by recycling a closed-loop coolant mixture of water and propylene glycol. The system can operate at temperatures as high as 115°F, allowing many facilities to avoid the water-intensive cooling towers and chillers traditionally used to prevent AI chips from overheating.

The announcement comes as scrutiny grows over the environmental impact of AI infrastructure and as major tech companies seek ways to curb water consumption and lower operating costs. However, NVIDIA acknowledged the benefits vary by climate, with hotter regions still likely to require supplemental cooling during extreme temperatures. The design also addresses only one aspect of AI’s environmental footprint, leaving broader concerns around energy consumption, fossil fuel use, and data center expansion unresolved. → Read the full article here.

NEWS

What Else is Happening

Google Backs A24 With $75M (Paywall): Google is investing about $75 million in A24 and launching an AI research partnership with the studio behind recent hits, marking the tech giant's first ownership stake in a film company.

Groq Raises $650M Post-Nvidia Not-Acqui-Hire: Six months after Nvidia licensed its tech and poached its CEO, Groq raised $650M to pivot from chipmaker to inference-focused neocloud under co-founder Doug Wightman.

SpaceX Taps Bond Market After IPO: Days after its IPO, SpaceX launched its first bond offering to refinance short-term debt and fund AI infrastructure and Starship development while avoiding shareholder dilution.

Chevron Powers Microsoft AI Campus: Chevron signed a 20-year deal to supply natural-gas power to Microsoft's Texas data center campus, supporting up to 2.67 gigawatts as AI-driven electricity demand surges.

POLL

Can We Afford to Slow AI Down?

That's All for Today

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team

Build the Loop, Not the Agent

Namrata Ganatra — Tue, 23 Jun 2026 00:08:04 +0000

_{Guest contributors:}_{Ethan Wang}_{is founding member of Google DeepMind Spark agent and founding member of Google DeepMind Mariner Agent.}_{Namrata Ganatra}_{is VP of Product & Engineering at Intuit, founder of AI-Native Builders, former AI startup founder, and ex-leader at Meta, Coinbase, Xero, and Microsoft.}

The better bet is simple: the team that can iterate fastest against the newest models will win.

Almost every architectural decision should be judged by one test — does it shorten the iteration loop or lengthen it? Three principles follow: make rapid iteration the primary objective, concentrate evaluation taste in a small elite team, and resist premature abstraction in favor of closing the end-to-end loop first.

1. Modernization Is a Continuous Process, Not a One-Time Project. Fund This Loop, Not the Artifact

If model capability keeps moving, your primary investment should not be any one modernized agent — it should be the infrastructure and practices that make each modernization cycle cheap and fast. When a new model arrives, the question you want to answer is “how quickly can we re-tune, re-evaluate, and ship?” — measured in days, not quarters.

What to fund: evaluation and CI infrastructure for agents, fast offline and online experiment tooling, and observability that makes failure modes legible. A useful rule of thumb is that the eval team should be at least as large as the agent-building team, because the loop — not the agent — is the asset, and the loop is gated by how fast and how trustworthy your evaluation is.

2. Concentrate Quality “Taste” in a Small, Elite Team

The hardest part of building great agents is not the plumbing — it is judgment. The ability to look at model output and know what is genuinely good for your customers, and to have an instinct for what the data is really saying, is a skill that compounds with repetition. That taste is the single most valuable AI asset a company can own, because it transfers across model generations even as the underlying technology churns.

This argues against spreading evaluation responsibility thinly across every product team. Talent this rare should be concentrated: put your best people on a small team whose full-time job is hill-climbing agent quality. They build the instinct; they own the bar. Diffuse the judgment across many teams and communication cost rises while the signal you get back fragments into inconsistent rubrics and thresholds.

The division of labor is clean. Product teams contribute the two things only they can — domain-specific evaluation sets and tool lists that encode what “good” means for their surface. The central team consumes those inputs and hill-climbs across all of them. This keeps domain knowledge where it lives while concentrating evaluation judgment where it compounds. Protecting this team from being diluted or reorganized into feature work is itself a strategic decision.

3. Defer Over-Engineering; Ship End-to-End First

A common over engineering example is to build an abstraction layer up front for future models — future-proofing against whatever comes next. In practice you cannot reliably predict the shape or scope of the changes coming, so abstractions designed today usually miss, and meanwhile you pay a real upfront tax in development overhead, slower iteration, and complexity for flexibility you may never use. Worse, elaborate non-AI plumbing has a way of draining the team's momentum away from the only thing that actually moves quality.

So get the feedback loop small and let the system evolve. Reach a working end-to-end path first, even an ugly one — a rough-but-complete pipeline teaches you far more, far faster, than a sophisticated system that is not yet whole. Optimize for learning velocity now; let generality earn its place later, when the recurring need for it is something you have observed rather than guessed.

Putting It Together

The three principles are one machine viewed from three angles.

Principle 1 builds the loop and makes it fast.
Principle 2 supplies the human judgment that makes the loop produce quality rather than just motion.
Principle 3 removes the architectural overhead that would otherwise slow the loop down.

Together they describe a single operating model: a small elite team hill-climbing quality, fed by product-team eval sets and tool filters, running on infrastructure built for speed, shipping end-to-end and deferring generality until it earns its place. That team — and the taste it accumulates — becomes the most valuable AI asset the organization owns, because it is the one thing that survives every model swap.

Ethan Wang is a member of Google DeepMind, focused on building general-purpose agents on the path to AGI. He is a founding member of two of the org's flagship agent projects: Gemini Spark, the recently released general agent; and Mariner, company's first computer use agent.

Namrata Ganatra is Product and Engineering Leader, VP at Intuit and founder of the AI-Native Builders community where she is also an AI instructor. She previously founded an AI e-commerce startup and has held senior product and engineering leadership roles at Meta, Coinbase, Xero, and Microsoft.

SpaceX's Cursor Deal, Anthropic's vs Washington, & China's Chip Battle

Matthew Berman — Mon, 22 Jun 2026 11:01:00 +0000

2026 State of AI in the Enterprise: Box’s Report Available Now!

Good morning. It's Monday, June 22, and we're covering SpaceX's move into AI coding software, Anthropic's clash with Washington, and the growing fight over China's chip ambitions.

Plus: a prompt that turns today's ambiguity into tomorrow's priorities.

YOUR DAILY ROLLUP

Building the Secure File System for Agents

The first phase of AI was defined by AI model access— phase two is defined by access to context.

96% of organizations know agents need access to their unique IP, the challenge now is connecting agents to trusted internal content.

Box’s State of AI report analyzes this adoption barrier and details how organizations can make enterprise knowledge accessible.

Download the report to learn more!

Your AI Roadmap, One Working Session

Tell us where you're stuck. We'll sit down with your team, find the highest-leverage use case, and leave you with a plan to ship it.

Book a session →

No cost. No pitch. No slide deck.

VIDEO

You Need to Try AI Loops

AI coding "loops" let agents work autonomously toward goals, automating tasks like testing, optimization, docs, bug fixes, and SEO.

REGULATION

Anthropic's AI Success Sparks High-Stakes Clash With Washington

Anthropic's rapid rise as an AI powerhouse is putting it on a collision course with the U.S. government, according to a report from The Economist. The AI company, which built its business around enterprise tools like Claude Code, is reportedly approaching a $1 trillion valuation and says it expects to become the first major AI lab to post an operating profit.

At the same time, tensions with the Trump administration have escalated over AI safety, national security, and access to Anthropic's most advanced models, leading to restrictions on some foreign users and raising questions about government oversight. The dispute underscores a broader challenge facing the AI industry: balancing commercial growth and innovation with growing demands for regulation and national security safeguards as frontier models become more powerful. → Read the full article here. (Paywall)

REINVENTION

Allbirds' AI Pivot Starts With a CEO and No Team

Former shoe brand Allbirds has completed its transformation into AI infrastructure company Smartbird, with new CEO Nadia Carlsten taking the helm as the business begins with no employees and no office. After selling its footwear business for $43 million and raising $100 million, Smartbird plans to build dedicated AI computing clusters for customers that prioritize data sovereignty and direct control over their infrastructure, such as pharmaceutical, energy, financial, and public-sector organizations.

Carlsten argues the company is targeting a niche underserved by hyperscale cloud providers, though established players like Hewlett Packard and Equinix already offer similar services. The unconventional pivot turned a meme-stock narrative into a well-funded startup experiment, and Smartbird now faces the harder task of proving there's a durable business behind the AI rebrand. → Read the full article here.

ACQUISITION

SpaceX Bets on Cursor to Strengthen Its AI Ambitions

SpaceX will acquire Cursor parent company Anysphere, marking the aerospace giant's first major acquisition as a public company. The reported $60 billion deal gives SpaceXAI — created after SpaceX's February acquisition of xAI — ownership of one of the fastest-growing AI coding platforms, adding a popular developer product to its AI portfolio.

Analysts say the move supports a vertical integration strategy, combining SpaceX's computing infrastructure, xAI's model development, and Cursor's widely adopted application layer. While the acquisition could help attract developers and better utilize AI computing resources, SpaceXAI still faces stiff competition from leaders including Anthropic, OpenAI, and Google and must improve its underlying AI models to close the gap. → Read the full article here.

NEWS

What Else is Happening

Amazon Probes AI Data Center Critics: Amazon is investigating three engineers who opposed AI data center expansion at Seattle hearings, sparking backlash.

ARD Aims to Standardize AI Discovery: ARD lets AI find agents across open catalogs, aiming to standardize discovery without replacing existing systems.

MIT Study Links AI Reliance to Weaker Judgment: Chatbots improved accuracy but heavy users got worse at spotting misinformation without AI.

General Intuition Eyes $300M Raise: The AI startup is seeking funding at a $2B valuation to scale agent training built on Medal’s massive gaming dataset.

Forward Future by Matthew Berman

On-Device AI, OpenAI Sol Safety, & GPT-5.6 Efficiency

YOUR DAILY ROLLUP

Top Stories of the Day

FRIDAY FACT

One Second of Your Mind

POWERED BY HIGGSFIELD AI

Claude Just Learned to Make Videos

VIDEO

GPT 5.6 Becomes a Workhorse

FORWARD FUTURE REVIEWS

Matt Spent 25 Billion Tokens on GPT-5.6 — Here's His Verdict

ON-DEVICE AI

PrismML Claims Record 27B-Parameter AI Model Running on iPhone

REGULATION

Questions Persist Over Government Approval of OpenAI's Sol AI Model

ZEITGEIST

The Week AI Became a Cost Question

NEWS

What Else is Happening

FRIDAY FACT

One Second of Brain Activity Took 40 Minutes To Simulate

TWEETS

NEO’s Hands To Tackle Human Tasks

POLL RESULTS

Does AI Make Planned Economies Viable?

That's All for Today

AI Search Debate, Infrastructure Debt Boom, & Grok 4.5

YOUR DAILY ROLLUP

Top Stories of the Day

POWERED BY HIGGSFIELD AI

Claude Just Learned to Make Videos

VIDEO

The Hidden Mind of AI

ZEITGEIST

When AI Knows Too Much

POLL

Does AI Make Planned Economies Viable?

CURIOSITY

AI Search May Be Making Us Less Curious, Neuroscientist Argues

INVESTMENT

AI Infrastructure Boom Fuels Record Corporate Bond Issuance

NEWS

What Else is Happening

POLL RESULTS: REAL OR AI

Can You Still Tell if an Image Is AI?

That's All for Today

Chinese AI Models, Illinois AI Law, & Alibaba Bans Anthropic

YOUR DAILY ROLLUP

Top Stories of the Day

VIDEO

Slash AI Costs With Model Routing

FORWARD FUTURE ORIGINAL

The New Coding Agents Problem: They Don’t Work Together

COMPETITION

Chinese AI Models Gain Ground as U.S. Companies Cut AI Costs

REGULATION

Illinois Enacts AI Safety Law Targeting Largest AI Model Developers

ROBOTICS

AI Advances Bring Autonomous Robots Closer to Everyday Work

NEWS

What Else is Happening

REAL OR AI

Can You Still Tell if an Image Is AI?

POLL RESULTS

Can AI Be Trusted Unsupervised?

That's All for Today

The New Coding Agents Problem: They Don’t Work Together

What Breaks When Agents Don’t All Come From the Same Place

What You’re Gaining From Mixed Fleets

AI Competition, Superintelligence, & the Future of Work

YOUR DAILY ROLLUP

Top Stories of the Day

POWERED BY BOX

Box Report: Governance Must Be Designed for AI

POLL

Can AI Be Trusted Unsupervised?

VIDEO

Codex Clones Excel Autonomously

GEOPOLITICS